Azure MLで試して学ぶAIによる自動走行とAIの仕組みハンズオン

Transcript

1. "[VSF.-Ͱࢼֶͯ͠Ϳ
   "*ʹΑΔࣗಈ૸ߦͱ"*ͷ࢓૊ΈϋϯζΦϯ
   "-(:"/Πϕϯτ
   !NBTBUP@LB ·͞ͱ͔
   
   

   View Slide

2. ຊ೔ͷλΠϜεέδϡʔϧ
   • AI RC CARʹ͍ͭͯ(30෼)
   • ϋϯζΦϯύʔτ(120)
   • ؀ڥߏங(20෼)
   • ԋश̍(40෼)
   • ԋश̎(60෼)
   

   View Slide

3. "*3$$"3ͱ͸ʁ
   ࢢൢϥδίϯ4#$
   "*ͱΧϝϥͷΈ
   -&7&-૬౰ͷ૸ߦ
   ੈքதͰϨʔε͕։࠵͞Ε͍ͯΔ
   ৄ͘͠͸IUUQNBTBUPLBIBUFOBCMPHDPN
   

   View Slide

4. ϋʔυ͔Β"*·ͰϑϧελοΫ
   ϋʔυ΢ΣΞ "*$7੍ޚ ύΠϩοτదੑ
   "*Λத৺ͱͨ͠ΤϯδχΞϦϯάͷूେ੒
   ৄ͘͠͸IUUQNBTBUPLBIBUFOBCMPHDPN
   

   View Slide

5. "*3$$BS૸ߦͷखॱ
   4FMG%SJWJOH
   $JSDVJU
   ………..
   ڭࣔσʔλͷऩू "*ͷֶश ਪ࿦૸ߦ
   ϧʔϧΛ࡞Βͣʹσʔλ͔ΒϧʔϧΛ࠶ݱ͢Δ
   ৄ͘͠͸IUUQNBTBUPLBIBUFOBCMPHDPN
   

   View Slide

6. ਤղ͜Ε͕"*3$$"3ͷ"*ͩʂ
   ը૾͔Βಛ௃తͳ෦෼Λநग़݁͠ՌΛܭࢉ͢Δ
   ৞ΈࠐΈ૚ $POW
   શ݁߹૚ 'VMMZ$POOFDUFE
   
   ೖྗ
   εςΞϦϯά
   ΞΫηϧ
   ը૾ͷಛ௃తͳ෦෼Λநग़ ಛ௃͔Β݁ՌΛٻΊΔ
   ݁Ռ
   

   View Slide

7. AIͷֶशͱ͸ʁ
   ৞ΈࠐΈ૚(Conv) શ݁߹૚(Fully-Connected)
   LOSS(໨తؔ਺ʣ
   015*.*;&3
   ೖྗ
   ͓खຊͱͷޡࠩΛ࠷খԽ͢ΔύϥϝʔλΛٻΊΔ
   ͓खຊ
   ݁Ռ
   ύϥϝʔλߋ৽
   ޡࠩͷܭࢉ
   

   View Slide

8. DENSE4
   DENSE3
   DENSE2
   DENSE1 FLAT CONV4
   CONV3
   ϏϧσΟϯάϒϩοΫͰߏ੒͞ΕΔAI
   ϥΠϒϥϦͷύʔπΛ૊Έ߹Θ࣮ͤͯݱ͢Δ
   CONV2
   LOSS
   OPTIMIZER
   CONV1
   

   View Slide

9. ϋϯζΦϯύʔτ
   

   View Slide

10. • 1. Azure Machine LearningΛ࢖ͬͨJupyter Notebook؀ڥͷߏங
    • 2. ԋश1. ڑ཭ηϯασʔλͱશ݁߹૚ʹΑΔࣗಈ૸ߦ
    • 3. ԋश2. CNNʹΑΔࣗಈ૸ߦͷֶश
    ϋϯζΦϯ
    

    View Slide

11. ౸ୡ໨ඪ
    ͱΓ͋͑ͣɺࣗ෼ͰAIͷϓϩάϥϜ
    Λॻ͚ͦ͏ͳؾ͕ͨ͠Βউͪɻ
    

    View Slide

12. 1. Azure Machine Learningͷߏங
    

    View Slide

13. Azure Machine Learningͷ࡞੒(1/2)
    1. Azure PortalͰAzure Machine LearningΛݕࡧ
    2.Azure Machine Learningͷ࡞੒ΛԡԼ
    

    View Slide

14. Azure Machine Learningͷ࡞੒(2/2)
    3.ϫʔΫεϖʔεͷ࡞੒
    ஫ҙϦʔδϣϯ͸ถࠃ౦෦ΛબͿ͜ͱ
    4.಺༰Λ֬ೝޙɺ࡞੒
    

    View Slide

15. Azure Machine Learning Studio
    1.σϓϩΠ׬ྃޙɺϦιʔεʹҠಈΛԡԼ
    2.Ϧιʔεը໘தԝԼଆʹදࣔ͞ΕΔϘλϯΛԡԼͯ͠
    Azure Machine Learning StudioΛ։͘
    

    View Slide

16. Notebooksͱԋश༻ϊʔτϒοΫͷ४උ
    1.Azure Machine Learning StudioΛ։͍ͨΒNotebooksͷʮࠓ͙͢։࢝ʯΛԡԼ
    2. Notebooksը໘ࠨଆͷϖΠϯΛ։͖ɺΞοϓϩʔυϘλϯ
    ͔Βԋश༻ͷϊʔτϒοΫΛΞοϓϩʔυ͢Δɻ
    https://drive.google.com/uc?export=download&id=1nEZFjS3gRR94vH7d0ZBa6tgdiaJsClV5
    ԋश༻ϊʔτϒοΫ:
    

    View Slide

17. ίϯϐϡʔςΟϯάϦιʔεͷ࡞੒
    1. GPUΠϯελϯεͷ্ཱͪ͛
    2. ΠϯελϯεͷىಈΛ଴ͭ
    3.Jupyter NotebookΛ։͘
    

    View Slide

18. • ফ͠ํ
    • ΠϯελϯεΛফ͢
    • MLΛফ͢
    • ϫʔΫεϖʔε
    ϋϯζΦϯऴྃޙͷΠϯελϯεͷ࡟আ
    ஫ҙɿϋϯζΦϯऴྃޙ͸ඞͣΠϯελϯεΛఀࢭɺ·ͨ͸࡟আ͠·͠ΐ͏ɻ
    

    View Slide

19. Πϯελϯεͷ࡟আ
    1. ࠨଆϝχϡʔͷʮ؅ཧʯ͔ΒίϯϐϡʔςΟϯάΛબ୒
    2. ࡞੒ͨ͠ΠϯελϯεʹνΣοΫΛೖΕͯʮ࡟আʯΛԡԼ
    

    View Slide

20. Azure ML Workspaceͷ࡟আ
    1. Azure ML Workspaceը໘Ͱʮ࡟আʯΛԡԼ͢Δɻ
    

    View Slide

21. Ϧιʔεάϧʔϓͷ࡟আ
    1.࡞੒ͨ͠ϦιʔεάϧʔϓΛબ୒͠ɺʮϦιʔεάϧʔϓͷ࡟আʯΛԡԼ͢Δ
    ɹ˞ ֬ೝͷͨΊɺ࡟আ͢ΔϦιʔεάϧʔϓ໊Λ໌ࣔతʹೖྗ͢Δɻ
    ผʹར༻͍ͯ͠ΔϦιʔεάϧʔϓͰ͋Ε͹ͦͷ··࢒͓ͯ͘͠ɻ
    

    View Slide

22. 2. શ݁߹૚ʹΑΔࣗಈ૸ߦ
    

    View Slide

23. ԋश1. શ݁߹૚ͷΈʹΑΔࣗಈ૸ߦ
    Input (5)
    Dense 16 ReLu
    Dense 16 ReLu
    output(2)
    ●Optimizer: Adam
    - lr: 0.001
    - Decay: 0.0
    ●Loss: MSE
    

    View Slide

24. جຊͷશ݁߹૚
    ઢܗɾඇઢܗͷม׵Λ܁Γฦ͠೚ҙͷؔ਺Λදݱ
    w11
    i1
    i2
    i3
    w21
    w31
    w12
    w22
    w32
    w13
    W23
    W33
    o1
    o2
    o3
    a1
    a2
    a3
    activation(o1)=a1
    activation(o2)=a2
    activation(o3)=a3
    ReLu
    ଟ૚ߏ଄ʹ͢Δ͜ͱͰ೚ҙͷؔ਺Λදݱ
    ֶशͰ͸8ͷ஋Λௐ੔͢Δ
    [i1
    i2
    i3]
    w11
    w12
    w13
    w21
    w22
    w23
    w31
    w32
    w33
    + b = [o1
    o2
    o3]
    iW + b = o O
    i
    

    View Slide

25. keras.layer.Dense
    https://www.tensorflow.org/api_docs/python/tf/keras/layers/Dense
    layers.Dense(16, activation='relu')
    layers.Dense(16, activation='relu', input_shape=(5,))
    Ϣχοτ਺ ΞΫςΟϕʔγϣϯؔ਺
    ೖྗ਺(NNͷೖྗͷΈࢦఆ)
    શ݁߹૚ͷఆٛ
    

    View Slide

26. ڭࢣσʔλͱͷޡࠩΛܭࢉ͢Δ໨తؔ਺
    "*ͷֶशํ޲ੑΛܾΊΔ
    .4& ฏۉೋ৐ޡࠩ
    $SPTT&OUSPQZޡࠩ
    શͯͷσʔλͱਪ࿦݁Ռͷޡࠩͷ૯࿨ͷೋ৐
    ʔ
    ̎
    AIͷਪ࿦݁Ռ͕͓खຊʹۙ͘ͳΕ͹஋͕খ͘͞ͳΔ
    ϥϕϧ͕ਖ਼ղͨ࣌͠ʹ஋͕খ͘͞ͳΔؔ਺
    ෼ྨ໰୊(ೣ͔ݘ͔ɺը૾ͷ਺ࣈ౰ͯ)
    0.7
    0.3
    0.1
    1.0
    0.0
    0.0
    

    View Slide

27. ύϥϝʔλΛௐ੔͢ΔOptimizer
    ࠷ٸ߱Լ๏Ͱޡ͕ࠩখ͘͞ͳΔΑ͏ʹύϥϝʔλΛߋ৽
    Loss
    Weight
    ޯ഑(ඍ෼)
    ٻΊ͍ͨ࠷খ஋
    ɾSGD
    ɾAdam
    ɾRMSProp
    Α͘࢖ΘΕΔख๏
    ॳظ஋-αXޯ഑
    ॳظ஋
    ऩଋ͢Δ·Ͱஞ࣍
    ܭࢉΛ܁Γฦ͢
    

    View Slide

28. ԋश̎. CNNʹΑΔࣗಈ૸ߦ
    Input 120x160x3
    kernel(5,5), 24, Stride(2,2), Relu
    kernel(5,5), 32, Stride(2,2), Relu
    kernel(5,5), 64, Stride(2,2), Relu
    kernel(3,3), 64, Stride(1,1), Relu
    kernel(3,3), 64, Stride(1,1), Relu
    Dense 100, Relu
    Dense 50, Relu
    Flatten
    output(2) ●Optimizer: Adam
    - lr: 0.001
    - Decay: 0.0
    ●Loss: MSE
    ૝ఆೖྗը૾(120x160x3)
    {"cam/image_array": “2584_cam-image_array_.jpg",
    "user/angle": 1.0, "user/throttle": 0.5,
    "user/mode": "user", "milliseconds": 272202}
    ڭࢣϥϕϧ
    

    View Slide

29. ը૾ͷಛ௃Λநग़͢Δ৞ΈࠐΈ૚
    Կ૚ʹ΋܁Γฦ͠ίʔεͷಛ௃Λֶश͢Δ
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    0.5 0.5 0.5
    ϑΟϧλ஋ͱ֤ըૉͷੵΛ଍͜͠Ή
    .BY1PPMJOH૚
    ֶशͰ͸ϑΟϧλͷॏΈΛௐ੔͢Δ
    

    View Slide

30. ৞ΈࠐΈ૚ͷύϥϝʔλͷߟ͑ํ
    ύϥϝʔλΛௐ੔͢Δ͜ͱͰग़ྗͷαΠζΛௐ੔
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    ೖྗ24x24x3(RGB) ϑΟϧλαΠζ3x3x2
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    W
    Stride:(1,1)
    ϑΟϧλΛಈ͔͢෯
    (24)x(24)x2=1152
    ࢀߟ:https://www.hellocybernetics.tech/entry/2016/12/23/000557
    Isize
    + 2P − Fsize
    Stride
    + 1 = Output
    Padding:(1,1)
    ग़ྗը૾ͷपΓΛຒΊΔ
    22 1
    1
    1
    1
    W
    W
    W
    W
    W
    W
    W
    W
    W
    Stride(1,1)
    ը૾ͷ୺·Ͱ
    24-3+1=22
    

    View Slide

31. Keras.layers.Conv2D
    layers.Conv2D(32, (5,5), strides =(2,2), activation="relu", name="conv2d_2"),
    layers.Conv2D(24, (5,5), strides =(2,2), input_shape=(120, 160, 3), activation="relu", name="conv2d_1"),
    https://www.tensorflow.org/api_docs/python/tf/keras/layers/Conv2D
    ϑΟϧλຕ਺ ϑΟϧλαΠζ
    ೖྗ਺ը૾αΠζ
    ετϥΠυ
    

    View Slide

32. Keras.layers.Dropout
    layers.Dropout(0.2)
    ֶश࣌ʹར༻͠ͳ͍ॏΈΛ࡞ΓաֶशΛ๷͙
    w11
    i1
    i2
    i3
    w21
    w31
    w12
    w22
    w32
    w13
    W23
    W33
    o1
    o2
    o3
    a1
    a2
    a3
    

    View Slide

33. Keras.layers.MaxPooling2D
    https://tensorflow.org/api_docs/python/tf/keras/layers/MaxPool2D
    ϑΟϧλ಺Ͱ࠷େͷ஋Λ࠾༻͢Δ
    W
    W
    W
    W
    W
    W
    W
    W
    W
    layers.MaxPool2D(pool_size=(3, 3),
    strides=(3,3), padding='valid')
    ϓʔϧαΠζ
    ετϥΠυ
    

    View Slide

34. ࣮͸/7*%*"ͷ࿦จͱ΄΅ಉ࣮͡૷
    BOJARSKI, M., DEL TESTA, D., DWORAKOWSKI, D., FIRNER, B., FLEPP, B., GOYAL, P., JACKEL, L. D., MONFORT, M., MULLER, U., ZHANG, J., ET AL. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)
    Dense 100, Relu
    Dense 50, Relu
    Flatten
    output(2)
    

    View Slide

35. /7*%*"ͷ࿦จʹཱͪ໭ͬͯΈ·͠ΐ͏
    BOJARSKI, M., DEL TESTA, D., DWORAKOWSKI, D., FIRNER, B., FLEPP, B., GOYAL, P., JACKEL, L. D., MONFORT, M., MULLER, U., ZHANG, J., ET AL. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)
    ೖྗը૾
    ̑૚ͷ৞ΈࠐΈ૚
    Ͱಛ௃நग़
    ̐૚ͷશ݁߹૚Ͱਓ
    ͷૢ࡞ϧʔϧΛ໛฿
    1࣍ݩ഑ྻʹม׵
    

    View Slide

36. ྑ͍AIΛ࡞Δ࠷େͷ
    ίπ͸ྑ͍σʔλΛ
    ूΊΔ͜ͱͰ͢ɻ
    

    View Slide

37. • શ݁߹૚ͱCNN
    • సҠֶश(෺ମݕग़ͳͲʣ
    • ϞσϧͷѹॖɺTensorFlow Lite, TensorFlow Lite for microcontroller
    • Edge TPUͱ͍ͬͨσΟʔϓϥʔχϯάΞΫηϥϨʔλɺϞόΠϧσόΠεͰͷਪ࿦
    • PytorchͳͲͷผͷϥΠϒϥϦΛ৮ͬͯΈΔɻ
    • RNN΍LSTM, Transformer
    • AutoEncoder, GANͳͲੜ੒ܥϞσϧ
    AI؆୯ͩͳͱࢥͬͨΒઌʹਐ΋͏
    ࠓ೔ཧղͨ͜͠ͱ͸શମ͔ΒݟͨΒ͘͝Ұ෦Ͱ͢
    

    View Slide

38. ·ͱΊ
    "*։ൃΛͰ͖ͦ͏ͳؾ͕͠·͔ͨ͠ʁ
    ਓͷӡసͷ໛฿ͷ࢓૊Έ
    "*։ൃ͸ϏϧσΟϯϒϩοΫ
    ͔͜͜ΒֶࣗࣗशΛ࢝ΊΑ͏
    ৄ͘͠͸IUUQNBTBUPLBIBUFOBCMPHDPN
    

    View Slide

39. https://drive.google.com/uc?export=download&id=1YfYzN6TOKUiobD7S-qgPblrRl2ieS4ba
    ϋϯζΦϯͷ׬੒൛ϊʔτϒοΫ
    

    View Slide

40. View Slide